Per @rauchg, 80% of token traffic on Vercel's AI gateway now goes to open models (up from 10% in December '25)

Some notes from @MarioMartone01 of Elise AI on building reliable AI products at scale (5M+ voice calls per month):

- team of 10 focused on evals / analyzing calls
- aim for 99.9%+ on reliability, every 9 is a hill to climb
- only look at models on internal benchmarks since performance can diverge from public ones

Claude "leads" 26% of Anthropic's AI R&D work as of September, up from under 1% in February.

Salesforce is acquiring Braze for ~$5.7B in cash. Some notes:

• $47.50/share, a 58% premium to trading price
• ~$911M in revenue expected this year growing 26%, so Salesforce paying ~6x revenue
• ~10% non-GAAP operating margin, still GAAP unprofitable
• ~2,770 customers incl. Canva, Instacart, Venmo and Dairy Queen

TSMC Capex by year

2023: $29.4B
2024: $29.7B
2025: $40.9B
2026: ~$62B

Interesting analysis by cursor breaking down agent inference spend.

They were able to reduce token use by trimming the system prompt, loading tools on demand and compressing file reads

cursor.com/blog/improved-…

Long term I think most brands and marketplaces will both play well with external agents and launch their own shopping agent for their storefront.

Levi's recently launched an AI shopping assistant on their website, and shoppers who use it add to cart at ~3x the avg rate

Good piece on how OpenAI thinks about its FDE efforts and how the role now involves making yourself obsolete every few months

openai.com/index/forward-…

Bank of America had "a couple hundred million" in additional expenses in Q3 on two things: AML compliance and Mythos remediation.

Anthropic's Mythos model exposed vulnerabilities in their code and they brought in 1,000 extra contractors to help patch them.

ASML will ship ~65 low-NA EUV systems in 2026 and plans 30% more capacity in 2027.

They say their increased "supply as currently planned for both 2027 and 2028 will not fully meet the demand" of customers

Per @hwchase17, Langchain spends ~30% of engineering salary on coding agents.

That percentage is down over the last few months while token usage went up, thanks to better visibility, cost controls and a better harness

TIL Nvidia has asked all three memory suppliers to build NVHBM base dies on TSMC's 3nm process.

It's the first time Samsung has worked with TSMC at the HBM stage, though they will retain in-house production for standard HBM

Anthropic has gone from ~2.5k code changes per day to 30K per day in 4 months

Great piece on their building of an internal version control system called Scion anthropic.com/engineering/re…

Interesting that Replit switched its default model to OpenAI. It was until recently a top-five Anthropic customer.

Amjad cited performance, economics, and the strength of their partnership with OpenAI as the reasons for the change.

Congrats to the Dream Labs team on the launch of Dream-1 which achieves SOTA while being faster and cheaper than other models. The model is available to try publicly today!

We @Wing_VC are thrilled to partner with the team in their quest to make generalized robotics intelligence accessible

Today, Dream Labs is coming out of stealth with Dream-1, a robot foundation model that controls bimanual robots it has never seen, on tasks it was never shown, in places it has never been, from $3/hour.

Dream API is live for everyone, today. 🧵

Per Alex Mashrabov of @higgsfield: internally ~400 people spend $4M+/month on models, over $10K per person per month.

I like the concept of token-market fit from @TaliaGold: categories where the average end user productively can spend >$10K/month in tokens.

Major categories today include coding, media/video, high frequency trading

Micron CEO Sanjay Mehrotra on memory supply and demand:

"Even with additional industry DRAM cleanroom space plans, we do not have line of sight to when supply and demand will return to balance."

He expects even tighter supply in 27/28 relative to demand than in '26

I like Turing CEO @Jonsid's framing on how Enterprise's should think about AI over time.

For non-core workflows, its okay to rent intelligence. For important core workflows, you want to own the learning loop: have custom evals, deploy a system, record traces, fine-tune & hill climb on accuracy/cost/latency